github_mirrors/vectorscan
3
0
Fork 1
mirror of https://github.com/VectorCamp/vectorscan.git synced 2025-09-30 03:34:25 +03:00
Code Issues Packages Projects Releases Wiki Activity

lookaround:

Browse Source 
add 64x8 and 64x16 shufti models
add mask64 model
expand entry quantity
This commit is contained in:
Hong, Yang A
2020-10-20 20:34:50 +00:00
committed by Konstantinos Margaritis
parent 56cb107005
commit dea7c4dc2e
12 changed files with 892 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
src/util/arch/x86/simd_utils.h
View File

@@ -424,6 +424,11 @@ static really_inline m256 loadu256(const void *ptr) {
return _mm256_loadu_si256((const m256 *)ptr);
}
static really_inline
m256 loadu_maskz_m256(__mmask32 k, const void *ptr) {
return _mm256_maskz_loadu_epi8(k, ptr);
}
// unaligned store
static really_inline void storeu256(void *ptr, m256 a) {
_mm256_storeu_si256((m256 *)ptr, a);
@@ -712,6 +717,22 @@ m512 loadu512(const void *ptr) {
return _mm512_loadu_si512(ptr);
}
// unaligned store
static really_inline
void storeu512(void *ptr, m512 a) {
#if defined(HAVE_AVX512)
_mm512_storeu_si512((m512 *)ptr, a);
#elif defined(HAVE_AVX2)
storeu256(ptr, a.lo);
storeu256((char *)ptr + 32, a.hi);
#else
storeu128(ptr, a.lo.lo);
storeu128((char *)ptr + 16, a.lo.hi);
storeu128((char *)ptr + 32, a.hi.lo);
storeu128((char *)ptr + 48, a.hi.hi);
#endif
}
static really_inline
m512 loadu_maskz_m512(__mmask64 k, const void *ptr) {
return _mm512_maskz_loadu_epi8(k, ptr);
@@ -722,6 +743,11 @@ m512 loadu_mask_m512(m512 src, __mmask64 k, const void *ptr) {
return _mm512_mask_loadu_epi8(src, k, ptr);
}
static really_inline
void storeu_mask_m512(void *ptr, __mmask64 k, m512 a) {
_mm512_mask_storeu_epi8(ptr, k, a);
}
static really_inline
m512 set_mask_m512(__mmask64 k) {
return _mm512_movm_epi8(k);

41
src/util/copybytes.h
View File

@@ -1,5 +1,5 @@
/*
* Copyright (c) 2016, Intel Corporation
* Copyright (c) 2016-2020, Intel Corporation
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
@@ -33,7 +33,7 @@
#include "simd_utils.h"
static really_inline
void copy_upto_32_bytes(u8 *dst, const u8 *src, unsigned int len) {
void copy_upto_64_bytes(u8 *dst, const u8 *src, unsigned int len) {
switch (len) {
case 0:
break;
@@ -72,14 +72,41 @@ void copy_upto_32_bytes(u8 *dst, const u8 *src, unsigned int len) {
case 16:
storeu128(dst, loadu128(src));
break;
case 32:
storeu256(dst, loadu256(src));
break;
default:
assert(len < 32);
case 17:
case 18:
case 19:
case 20:
case 21:
case 22:
case 23:
case 24:
case 25:
case 26:
case 27:
case 28:
case 29:
case 30:
case 31:
storeu128(dst + len - 16, loadu128(src + len - 16));
storeu128(dst, loadu128(src));
break;
case 32:
storeu256(dst, loadu256(src));
break;
#ifdef HAVE_AVX512
case 64:
storebytes512(dst, loadu512(src), 64);
break;
default:
assert(len < 64);
u64a k = (1ULL << len) - 1;
storeu_mask_m512(dst, k, loadu_maskz_m512(k, src));
break;
#else
default:
assert(0);
break;
#endif
}
}